Push Door ACT

LeRobot ACT policy trained on jiangjy0508/push_door, a 40-episode LeRobotDataset v3.0 dataset for pushing a lightweight stall door with a RealMan ECO63F arm.

Revisions

Deployable checkpoints are available as tags:

005000, 010000, 015000, 020000, 025000, 030000, 035000, 040000, 045000, 050000.

The repository main revision contains checkpoint 050000. Offline held-out loss was lowest at 005000; select the final policy using real-robot closed-loop evaluation rather than training loss alone.

lerobot-rollout \
  --policy.path=jiangjy0508/push_door_act \
  --policy.pretrained_revision=005000 \
  --robot.type=realman_eco63f \
  --task="Push the door open" \
  --fps=30 \
  --device=cuda

realman_eco63f is custom hardware and requires a separately installed LeRobot Robot plugin.

Model contract

  • Images: cam_top, cam_wrist, cam_front, RGB 640×480
  • State: 17 dimensions
  • Force/torque: 6 dimensions
  • Action: 7 dimensions, end-effector delta plus gripper
  • Control frequency: 30 Hz
  • ACT chunk size: 40; execution horizon: 10

Load the complete revision, including policy preprocessor and postprocessor files. Do not load model.safetensors alone.

Limitations

The gripper state and action are constant zero in the training dataset, so this model cannot operate a gripper. It is intended only for the direct push-door branch. Hardware safety filtering, force limits, workspace limits, collision detection, and emergency stop handling are required outside the learned policy.

Downloads last month
6
Safetensors
Model size
51.6M params
Tensor type
F32
·
Video Preview
loading

Dataset used to train jiangjy0508/push_door_act